Predictive models composed by radiomic features extracted from multi-detector computed tomography images for predicting low- and high- grade clear cell renal cell carcinoma: A STARD-compliant article.

To evaluate the values of conventional image features (CIFs) and radiomic features (RFs) extracted from multi-detector computed tomography (MDCT) images for predicting low- and high-grade clear cell renal cell carcinoma (ccRCC).Two hundred twenty-seven patients with ccRCC were retrospectively recruited. Five hundred seventy features including 14 CIFs and 556 RFs were extracted from MDCT images of each ccRCC. The CIFs were extracted manually and RFs by the free software-MaZda. Least absolute shrinkage and selection operator (Lasso) was applied to shrink the high-dimensional data set and select the features. Five predictive models for predicting low- and high-grade ccRCC were constructed by the selected CIFs and RFs. The 5 models were as follows: model of minimum mean squared error (minMSE) of CIFs (CIF-minMSE), minMSE of cortico-medullary phase (CMP) of kidney (CMP-minMSE), minMSE of parenchyma phase (PP) of kidney (PP-minMSE), the combined model of CIF-minMSE and CMP-minMSE (CIF-CMP-minMSE), and the combined model of CIF-minMSE and PP-minMSE (CIF-PP-minMSE). The Lasso regression equation of each model was constructed, and the predictive values were calculated. The receiver operating characteristic (ROC) curves of predictive values of the 5 models were drawn by SPSS19.0, and the areas under the curves (AUCs) were calculated.According to Lasso regression, 12, 19 and 10 features were respectively selected from the CIFs, RFs of CMP image and that of PP images to construct the 5 predictive models. The models ordered by their AUCs from large to small were CIF-CMP-minMSE (AUC: 0.986), CIF-PP-minMSE (AUC: 0.981), CIF-minMSE (AUC: 0.980), CMP-minMSE (AUC: 0.975), and PP-minMSE (AUC: 0.963). The maximum diameter of the largest axial section of ccRCC had a maximum weight in predicting the grade of ccRCC among all the features, and its cutoff value was 6.15 cm with a sensitivity of 0.901, a specificity of 0.963, and an AUC of 0.975.When combined with CIFs, RFs extracted from MDCT images contributed to the larger AUC of the predictive model, but were less valuable than CIFs when used alone. The CIF-CMP-minMSE was the optimal predictive model. The maximum diameter of the largest axial section of ccRCC had the largest weight in all features.

Introduction

Renal cell carcinoma (RCC) is the most common malignant tumor among the diseases of urinary system, accounting for about 85% of renal malignant tumors and 3% of body tumors. Clear cell renal cell carcinoma (ccRCC) is the most common subtype of RCC, accounting for about 70 to 80% of all kidney cancers.[ The latest WHO/ISUP (World Health Organization/International Society of Urological Pathology) grading system of RCC, with 4 stratifications (grade 1–4), is commonly used for evaluating the outcome of ccRCC and papillary RCC.[ Many studies found that the WHO/ISUP grade of ccRCC was closely correlation with the cancer specific survival (CSS) of the patient and drew a conclusion that the higher WHO/ISUP grade the lower CSS could be got.[ Kuthi et al[ found the 5-year survival rate of grade 1 or 2 ccRCC was significant better than that of grade 3 or 4. However, there were not significant differences between the 5-year CSSs of grade 1 vs grade 2 tumors (P=.550), and that of grade 3 vs grade 4 tumors (P = .226). They also found when the grade 1 and 2 were combined into low-grade group, and the grade 3 and 4 into high-grade group, the 5-year CSSs of the 2 groups revealed a significant difference (P < .0001). The similar results could also be found in some other studies.[ Preoperative needle biopsy and pathological section of the nidus are the most authoritative methods to determine the WHO/ISUP grade of ccRCC. However, their applications are limited due to their invasiveness. So, we aim to find if there is a noninvasive way to predict the pathological grade of ccRCC.

With the rapid development of medical imaging technology and equipment, especially the application of 64 or more rows multi-detector computed tomography (MDCT), the CT images with high isotropic resolution can be achieved, more precisely presenting the internal structure of lesion. Normally, we can perceive the information about size, shape, density, necrosis and blood supply of the lesions by naked eyes, but cannot for the RFs concealed in the MDCT images. The RF is a kind of feature reflecting the homogeneous phenomenon about the pixels in an image and can be quantitative analyzed by some special software.[ Since ‘Beyond Imaging’ was selected as the theme of the 2016 annual conference of the Radiological Society of North America (RSNA), the radiomics has become a buzzword now. It refers to the high-throughput extracting of massive imaging features describing the characteristics of the pixels of nidus in digital medical images.[ In a short time, a large number of studies about radiomics emerged, mainly focusing on preoperative accurate diagnosis, classifying, and prognosis of tumors.[

Owing to obvious characteristics on enhanced CT images, such as fast in and out style of enhancement in artery phase and polycystic necrosis, ccRCC can be differentiated from the other subtypes of RCC, such as papillary RCC, chromophobe RCC, and oncocytoma, without much difficulty.[ With the development of precision medicine and personalized treatment, more and more clinicians want to know the WHO/ISUP grade of ccCC before treatment or operation, in order to determine the way of therapy and roughly predict the cancer-free survival (CSS) if surgery. Though, there were some radiomic studies about ccRCC could be found, however, their RFs were drown mainly from gray level histograms (GLH) and the outdated Fuhrman grading system of RCC was referred.[ As far as we know, there are no radioimc studies about exploring the correlations between numerous RFs and WHO/ISUP grades of ccRCC. So, we conduct this study with an intention to uncover the relationships between the RFs and the WHO/ISUP grades of ccRCC.

The whole procedure of this research included 4 parts: extracting massive RFs from MDCT images; selecting the most predictive RFs by Lasso regression and then constructing 5 predictive models; using the predictive models to calculate the predictive value for each ccRCC; and drawing ROC curves of the predictive values, calculating the AUC and the cutoff value of each model. Our ultimate aim is to find the optimal predictive model composed by CIFs and RFs for recognizing low-grade (WHO/ISUP grade 1 and 2) and high-grade (WHO/ISUP grade 3 and 4) ccRCC.

Materials and methods

Patients

This study was a retrospective case-control investigation, approved by the Ethics Committee of Sichuan University, and the consents from patients were waived. Two hundred seventy-one patients with ccRCC confirmed by 2 pathologists through biopsy or surgical resection were searched from the picture archiving and communication system (PACS) of our hospital from January 2012 to June 2017. The inclusion criteria of cases were as follows: no previous treatment in the patients before CT examination; the MDCT scanning of the patient must comprise 3 phases [the non-contrast phase (NCP), the cortico-medullary phase (CMP), and the parenchyma phase (PP) of enhanced CT scanning]; every case must has an explicit WHO/ISUP grade confirmed by pathology. Overall, in this study, 7 cases were excluded for having previous treatments, 16 cases were excluded for inadequate phases of MDCT scanning, and 21 cases were excluded for non-explicit WHO/ISUP grades. Finally, 227 patients remained in this study, of which 130 cases were male patients, 97 cases were female patients, the age range was 10 to 85 years old, the mean age was 57.3 ± 12.9 years, 136 cases were low-grade ccRCCs (WHO/ISUP grades 1 and 2), and 91 cases were high-grade ccRCCs (WHO/ISUP grades 3 and 4) (Table 1).

Table 1

Demographic and clinical characteristics of patients with clear cell renal cell carcinoma (ccRCC).

CT acquisition

Two hundred and four patients were examined by a 64-multidetector spiral CT scanner (LightSpeed VCT, GE Healthcare), and 23 patients were examined by a 256-multidetector spiral CT scanner (Brilliance iCT, Philips Healthcare). The NCP, CMP, and PP were recorded in each patient. The scanning protocols of the 3 phases were as follows: using spiral scanning and thin-slice reconstruction in NCP, CMP, and PP; the scanning of CMP and PP were started at 30th and 90th second after the contrast agent was injected into the antecubital vein, respectively. Contrast agent (70–100 mL; Iopamiro, Bracco, Italy; Ultravist, Bayer, Germany) was injected into the antecubital vein by high pressure injector at a rate of 3.5 ml/s. The CT scanning parameters of the 3 phases were as follows: the tube voltage was 120 KV; the automated tube-current modulation based on weight was used; collimation width was 0.625 mm; the pitch was 0.967:1; the reconstruction thickness was 1.5 to 2.5 mm; the field of view (FOV) was 360 mm; and the matrix was 512 × 512.

Conventional image features (CIFs) acquisition

The CIFs were extracted from MDCT images in the PACS (Huahai Health, China) of our hospital by 2 radiologists with 10 and 15 years of experience in diagnosis of abdominal imaging, respectively, both blinded to the explicit WHO/ISUP grade of each ccRCC. The region of interest (ROI) was determined according to the consents of the 2 radiologists. When measuring the attenuation of the parenchyma of ccRCC, they put 3 ROIs with the areas from 2.0 to 5.0 cm2, and then took the mean attenuation of the 3 ROIs for study. When measuring the whole attenuation of the maximum axial section of ccRCC including the necrotic area, the contour of the ROI was recessed approximately 2 mm from the tumor margin to reduce the partial volume effect. In this study, the CIFs defined and consented by the 2 radiologists were presented in Table 2.

Table 2

List of the conventional image features (CIFs) extracted from multi-detector computed tomography (MDCT) images of each clear cell renal cell carcinoma (ccRCC).

Radiomic feature (RF) extraction (Fig. 1)

Image selection: the images of the maximum axial section of each ccRCC on CMP and PP were chosen. All the images extracted from PACS were anonymized. The reconstructive thickness of each image was 1.5 to 2.5 mm. The window width and window level of every MDCT image ranged from 300 to 400 HU and 45 to 65 HU, respectively.

Flowchart of radiomic features (RFs) extraction from MDCT image. (A) The image of the maximum axial section of clear cell renal cell carcinoma (ccRCC) on cortico-medullary phase of kidney (CMP) was chosen. (B) The region of interest (ROI) of tumor was drown by pink color on MaZda (version 4.6). (C) The gray level histogram (GLH) of tumor was calculated. (D) The gray level co-occurrence matrix (GLCM) of tumor was calculated. (E) The gray level run long matrix (GLRLM) of tumor was calculated. (F) The histogram of oriented gradient (HOG) of tumor was calculated. (G) The wavelet transformation (WT) of tumor was calculated. (H) The auto-regressive model (ARM) of tumor was calculated. (I) The high dimensional radiomic data set of tumor was constructed. ARM = auto-regressive model, CMP = cortico-medullary phase of kidney, ccRCC = clear cell renal cell carcinoma, GLCM = gray level co-occurrence matrix, GLH = gray level histogram, GLRLM = gray level run long matrix, HOG = histogram of oriented gradient, MDCT = multi-detector computed tomography, RFs = radiomic features, ROI = region of interest, WT = wavelet transformation.

The ROI definition: The irregular contour of ROI drew by manually was recessed approximately 2 mm from the margin of tumor to reduce the interference by adjacent tissue, such as fat and normal renal tissue.

The RFs extraction: The RFs were extracted by a free software package, MaZda (version 4.6)[ and primarily came from gray level histogram (GLH), gray level co-occurrence matrix (GLCM), gray level run long matrix (GLRLM), histogram of oriented gradient (HOG), wavelet transformation (WT), and auto-regressive model (ARM) of the MDCT images of each ccRCC. The RFs coming from GLH indicated the intensity distribution of image pixels, but without spatial information. The RFs deriving from GLCM indicated the changes of pixel signal intensities with the increase of pixel length and described the distributions of the pixels with the same gray level at different orientation. The RFs derived from GLRLM reflected the homogeneity of pixel gray levels in different directions. The RFs derived from HOG primarily reflected the contour information of the images. The RFs derived from WT indicated the frequencies of a certain cohort of pixel gray levels within images. The RFs deriving from ARM, in which the gray level of some pixel could be estimated by the neighboring pixels, manifested the homogeneity of image in some degree.[ Overall, from the MDCT images of CMP and PP of each ccRCC, 556 RFs were extracted from the ROI of nidus in the maximum axial section.

Data standardization

To minimize the effects of variations of contrast and brightness of images, the gray level values of pixels were normalized by putting into the range of μ ± 3σ (μ, mean gray level value; σ, standard deviation) before statistical analysis. The values of different kinds of RFs were normalized into the range (0, 1) by sigmoid function, respectively.

RFs screening and predictive model construction

The most predictive features among the primary dataset were selected by the least absolute shrinkage and selection operator (Lasso) regression, which is very suitable for the shrinkage of a high-dimensional dataset.[ By means of Lasso regression, the minimum mean squared error model (minMSE) composed by the most predictive features was obtained from the dataset of CIFs, RFs of CMP, and that of PP, respectively. The 3 minMSEs were abbreviated as CIF-minMSE, CMP-minMSE, and PP-minMSE, respectively. The combined model of CIF-minMSE and CMP-minMSE was abbreviated as CIF-CMP-minMSE, and that of CIF-minMSE and PP-minMSE was abbreviated as CIF-PP-minMSE.

Statistical analysis

The Lasso regression equation of each model for predicting low- and high-grade ccRCC was constructed, and the predictive value of each ccRCC (‘1.0’ defined as the standard value of low-grade ccRCC and ‘2.0’ defined as the standard value of high-grade ccRCC) was calculated. The receiver operating characteristic (ROC) curve of predictive values of each model was drawn by SPSS19.0 (SPSS Inc., Chicago, IL), and the area under curve (AUC) was calculated.

Results

Twelve features were selected from CIFs by Lasso regression and constructed the model of minMSE. The equation of Lasso regression of CIF-minMSE was as follows: CIF-minMSE_Y = 0.1741∗X1-0.0717∗X3-0.1379∗X4-0.2986∗X5+0.8157∗X6-0.6586∗X8+2.4304∗X9-0.5833∗X10-0.0103∗X11-0.0176∗X12-0.3042∗X13-0.2645∗X14 (the ‘∗’ in the equation means ‘multiply’).

Nineteen features were selected from the 278 RFs extracted from CMP by Lasso regression and constructed the model of minMSE. The equation of Lasso regression of CMP-minMSE was as follows: CMP-minMSE_Y = 1.0192∗Area-0.2748∗MinNorm+0.0493∗Hist_Skewness-0.2055∗Hist_Perc.50%+0.2595∗S(0,1)Entropy-0.1400∗S(0,2)SumOfSqs+0.0818∗S(2,2)SumAverg + 0.2582∗S(3,3)Entropy+0.0818∗S(4,0)SumEntrp+0.3582∗S(5,0)InvDfMom-0.0072∗S(5,-5)SumAverg +0.4781∗Vertl_RLNonUni+0.0187∗GrSkewness+0.0373∗GrKurtosis+0.1331∗Teta2-0.1561∗WavEnHL_s-1+0.2385∗WavEnLL_s-3-0.0505∗WavEnHH_s-3-0.1909∗WavEnLL_s-4.

Ten features were selected from the 278 RFs extracted from PP by Lasso regression and constructed the minMSE model. The equation of Lasso regression of PP-minMSE was as follows: PP-minMSE_Y=0.1597∗Area -0.3982∗MinNorm-0.1115∗S(5,5)DifVarnc+0.1722∗S(5,-5)Entropy-0.2169∗S(5,-5)DifEntrp +1.4737∗Vertl_RLNonUni+0.0308∗Teta2-0.0228∗WavEnHL_s-2-0.0180∗WavEnLH_s-3-0.0216∗WavEnHL_s-3.

In the Lasso equations of the models of CIF-minMSE, CMP-minMSE, and the PP-minMSE, we found the X1, X6, and X9 of CIFs, the Area, Skewness, S(0,1)Entropy, S(2,2)SumAverg, S(3,3)Entropy, S(4,0)SumEntrp, S(5,0)InvDfMom, Vertl_RLNonUni, GrSkewness, GrKurtosis,Teta2, and WavEnLL_s-3 of RFs from CMP, and the Area, S(5,-5)Entropy, Vertl_RLNonUni, and Teta2 of RFs from PP had positive weight coefficients. However, the X3, X4, X5, X8, X10, X11, X12, X13, and X14 of CIFs, the MinNorm, Hist_Perc.50%, S(0,2)SumOfSqs, S(5,-5)SumAverg, WavEnHL_s-1, WavEnHH_s-3, and WavEnLL_s-4 of RFs from CMP, and the MinNorm, S(5,5)DifVarnc, S(5,-5)DifEntrp, WavEnHL_s-2,WavEnLH_s-3, and WavEnHL_s-3 of RFs from PP had negative weight coefficients. The definitions and interpretations of the RFs selected by Lasso regression and the groups that they belonged to was shown in Table 3.

Table 3

The definitions and interpretations of the radiomic features selected by Lasso regression from cortico-medullary phase of kidney (CMP) or parenchyma phase of kidney (PP).

The scattergram and histogram the predictive values of CIF-CMP-minMSE, for example, for predicting low- and high-grade ccRCC were shown in Figures 2 and 3, respectively. The cutoff values, sensitivities, specificities, and AUCs of the 5 predictive models (CIF-minMSE, CMP-minMSE, PP-minMSE, CIF-CMP-minMSE, and CIF-PP-minMSE) were shown in Table 4. The ROC curves of the 5 models were shown in Figure 4. In addition, the largest diameter of the maximum axial section of ccRCC (the X9 of CIFs) gained also good performance in predicting low- and high-grade ccRCC individually. Its cutoff value was 6.15 cm, with a sensitivity of 0.901, a specificity of 0.963, and an AUC of 0.975.

Figure 2

The scatter diagram of the predictive values of CIF-CMP-minMSE for predicting low- and high- grade ccRCC. ccRCC = clear cell renal cell carcinoma, CIF-CMP-minMSE = combined model of CIF-minMSE and CMP-minMSE.

Table 4

The results of the 5 predictive models in predicting the low- and high-grade of clear cell renal cell carcinoma (ccRCC).

Figure 4

The centralized display of the ROC curves of the 5 predictive models. ROC = receiver operating characteristic.

The histogram of the predictive values of CIF-CMP-minMSE for predicting low- and high- grade ccRCC. ccRCC = clear cell renal cell carcinoma, CIF-CMP-minMSE = combined model of CIF-minMSE and CMP-minMSE.

Discussion

In this study, 14 CIFs and 556 RFs were extracted from the 3-phase MDCT images of each ccRCC. Using the features selected from CIFs and RFs by Lasso regression, we got predictive equation of CIF-minMSE, CMP-minMSE, and PP-minMSE, respectively, in which each predictive feature had its own weight coefficient positive or negative. We proposed 5 predictive models composed by the selected CIFs and RFs to differentiate low-grade (WHO/ISUP grade 1 and 2) from high-grade (WHO/ISUP grade 3 and 4) ccRCC. On account of the AUC from large to small, the order of the 5 predictive models was CIF-CMP-minMSE (AUC: 0.986), CIF-PP-minMSE (AUC: 0.981), CIF-minMSE (AUC: 0.980), CMP-minMSE (AUC: 0.975), and PP-minMSE (AUC: 0.963). It was found that the RFs either extracted from CMP or PP, especially the former, were able to improve the accuracy rate of predicting the low- and high-grade ccRCC and the optimal predictive model was CIF-CMP-minMSE. Furthermore, among all the features, the feature X9—the maximum diameter of the largest axial section of ccRCC—was found having the maximum positive weight in predicting the WHO/ISUP grade of ccRCC.

Some previous studies had indicated the values of RFs extracted from CT or MRI images in the diagnoses, classifications, prognoses, and identifications of many tumors, such as the cancers of lung, rectum, breast, stomach, and brain.[ The reasons speculated by researchers might be that the RFs were closely related to the internal structures of niduses and partly reflected the heterogeneities of tumors. About ccRCC, there were also some studies on RFs could be found. Kierans et al[ found that high-grade ccRCC had significantly higher skewness of GLH and correlations of GLCM on preoperative magnetic resonance (MR) diffusion-weighted imaging (DWI) than that of low-grade ccRCC, respectively. Lubner et al[ studied 157 patients with RCC and drew a conclusion that the entropy, mean of positive pixels, and standard deviation of the pixel distribution of GLH on CT images were associated with tumor histologic types, nuclear grades, and outcomes. Ding et al[ calculated texture-score using 4 RFs (RunLengthNonuniformity, Contrast, GrayLevelNonuniformity and 0.025Quantile) selected from 184 RFs extracted from MDCT images of each ccRCC and found the texture-score was valuable in the preoperative discrimination of the low from high grade ccRCC. Compared to the previous researches, some similar results could be got by this study, however, the larger numbers of patients (227 cases) and RFs (556 features) were studied and the outdated Fuhrman grade system of ccRCC was substituted by WHO/ISUP grade system adopted by ISUP after March 2012. Among the AUCs of the 5 models, we found the AUC of CIF-CMP-minMSE was the largest one and that of CIF-PP-minMSE was the second. That was to say, when combined with CIF-minMSE, the RFs of MDCT images were conducive to the higher rate of correctly recognizing the low and high grade ccRCC, and moreover, the RFs of CMP-minMSE were more valuable than that of PP-minMSE. However, we also found, when used alone, the AUC of CIF-minMSE was larger than that of CMP-minMSE or PP-minMSE. That was to say, the CIFs were more valuable than RFs of CMP or PP in predicting the grade of ccRCC individually.

Among all the features in this study, the maximum diameter of the largest axial section of ccRCC was found having the maximum positive weight in predicting the grade of ccRCC and its cutoff value was 6.15 cm, with a sensitivity of 0.901, a specificity of 0.963, and an AUC of 0.975. The similar results could also be found in some other studies. In Kierans’ study,[ the means of the tumor size of low and high grade ccRCC were 3.49 ± 1.57 cm and 6.94 ± 2.93 cm, respectively, and their difference was statistical significant. Frank and Zhang et al[ also found the tumor size of ccRCC was positively correlated with the nuclear grade of tumor, with 1 cm increaser in diameter of tumor the possibility of high grade increased by 32%. The reason why the significant difference appeared between the tumor sizes of low- and high-grade ccRCC is not confirmed, the speculated reason needing further prove may be that the tumor cells will become more and more malignant with the tumor growth.

There were several limitations in this study. First, this study was a single-center retrospective research, and the results need further validation from multi-center samples. Second, due to the different shape of each patient, the tube voltage and milliampere second of each CT examination were not identically. And, in all patients, 204 cases were examined by a 64 multi-detector CT scanner, the remained 23 cases were examined by a 256-multidetector CT scanner. Because of this, the measurement errors might increase during extracting RFs. Third, the RFs extracted from the maximum axial section of tumors did not entirely reflect the texture structures of the 3-dimensional tumor, which should be further studied in the future. Fourth, there was no independent validation cohort in this study to evaluate the overfitting of these models. Fifth, the sample size of this study is still some small, and more cases need to be added in the future.

Conclusions

In this study, we presented the whole procedure of the analysis about massive RFs. According to Lasso regression, we constructed 5 predictive models for discerning low- and high-grade ccRCC and proved the model of CIF-CMP-minMSE was the optimal predictive model. When combined with CIFs, the RFs extracted from MDCT images could contribute to the larger AUC of the predictive model. But, when used alone, the CIFs were more valuable than RFs. Nevertheless, among all the features, the CIF feature X9—the largest diameter of the maximum axial section of ccRCC—had the largest positive weight in predicting the grades of ccRCC.

Acknowledgments

This work was supported by the Science and Technology Bureau of Luzhou China (Grant number: 12126).

Author contributions

Conceptualization: Xiaopeng He, Hanmei Zhang, Fugang Han, Bin Song.

Data curation: Xiaopeng He, Tong Zhang.

Formal analysis: Hanmei Zhang, Tong Zhang.

Methodology: Hanmei Zhang, Fugang Han, Bin Song.

Project administration: Xiaopeng He, Bin Song.

Supervision: Bin Song.

Validation: Tong Zhang.

Writing – original draft: Xiaopeng He.

Writing – review & editing: Xiaopeng He, Bin Song.